Albedo-Watch runs entirely on Cloud Run: one service in asia-south1 (Mumbai), close to the people in India it’s built for. One container for everything. A two-stage Dockerfile builds the React and CesiumJS front end with Node, then copies it into a Python 3.12 image running FastAPI. The same container serves the web app, the API and the Gemini agents, so there’s one thing to deploy and one URL to share. Deploying from source. gcloud run deploy --source . hands the build to Cloud Build and stores the image in Artifact Registry. In the final week before the hackathon deadline I shipped more than 20 revisions this way, often several a day. Settings that fit the app: CPU stays on between requests, so background jobs keep running: refreshing NASA fire data, retraining the federated forecast models and pulling news photos. Startup CPU boost shortens cold starts. It scales from zero, with a cap of two instances. The Copilot’s longer agent calls (up to about 100 seconds) fit within Cloud R

Albedo Watch

An AI operating layer for clean air

An AI operating layer for clean air